Prompt Guard: Stop Injection, Jailbreak, and Prompt Leakage

Prompt Guard is BeyondGuard’s first line of defense at the model surface. It inspects every prompt — and every document the model reads — before inference, classifying adversarial intent and masking sensitive data on the way in and the way out.

​

What Prompt Guard protects

The model surface: the prompts users and applications send, the system prompts that define the model’s behavior, and the responses that come back. Because detection runs outside the model, enforcement holds even when the model is jailbroken, replaced, or tricked.

​

Threats it stops

• Prompt injection & jailbreak — instruction override, persona hijacking, encoded and obfuscated payloads, tag-wrapped injection, and multi-step DAN-style attacks (BG-01, BG-05).
• Safety-prompt manipulation — attempts to weaken policy through “neutral” or “safety” prompt injection (BG-15), mediated by the Safety Wrapper (BG-13).
• System prompt leakage — extraction attempts defeated by prompt sealing and integrity fingerprinting (BG-14).
• Sensitive-information disclosure — PII, PHI, and secrets masked inline on input and output (BG-02, BG-20, BG-24).

​

How it works

Every prompt is normalized — encodings, homoglyphs, zero-width characters — and classified by a 12B-parameter security model trained on 717K+ adversarial samples, semantically across 140+ languages, in under 300 ms. Each interaction resolves to allow, deny, mask, or rewrite with a confidence score and reason code, and the full event streams to your SIEM.

​

Controls

Prompt Guard runs BG-01, BG-02, BG-05, BG-13, BG-14, BG-15, BG-20, and BG-24. See the Controls Catalog for each control’s definition, scope, and benchmark.

​

OWASP coverage

Addresses LLM01 Prompt Injection, LLM02 Sensitive Information Disclosure, and LLM07 System Prompt Leakage from the OWASP LLM Top 10.

​

Related